A Foundation Model for Capturing Complexity of Menstrual Health Data.

Area of Science:

• Computational health research
• Women's health
• Artificial intelligence applications

Background:

• Menstrual cycle data is complex, variable, and often limited, hindering computational health research.
• Generative artificial intelligence (AI) presents a novel opportunity for modeling large-scale menstrual health datasets.

Purpose of the Study:

• To introduce and evaluate a generative foundation model for menstrual health data.
• To assess the model's ability to generate physiologically plausible synthetic cycles and realistic tracking behaviors.
• To examine the model's learned representations for temporal and symptomatic patterns and evaluate privacy risks.

Main Methods:

• Trained a generative foundation model on self-tracked menstrual data from over 1.2 million app users.
• Assessed synthetic data fidelity, realism of tracking behaviors, and privacy implications.
• Evaluated learned representations on downstream forecasting tasks.

Main Results:

• The generative AI model produced high-fidelity synthetic menstrual data closely mirroring real-world user data.
• No evidence of data leakage was found, indicating strong privacy preservation.
• Learned representations significantly outperformed baseline methods in forecasting tasks.

Conclusions:

• Generative AI holds significant potential for advancing menstrual health forecasting and enabling privacy-sensitive data sharing.
• This approach can facilitate scientific inquiry and improve women's health research.
• The developed model offers a robust tool for generating realistic synthetic menstrual data.